Lightweight Deep Network for the Classification of Breast Cancer Histopathological Images

Out of all cancers among female patients, breast cancer is on top in terms of prevalence. Many deep learning approaches have been investigated for the classification of histopathological images of breast cancer, however most tend to be either too complex or too large to adopt at a clinical level. In this study we propose a lightweight network, based on knowledge distillation, which performs almost the same as the teacher network but with significantly fewer parameters. Our student network consists of inverted residual blocks of MobileNetV2 and the ghost module. We performed our experiments on BreakHis, BACH and Kaggle breast cancer histopathological imaging datasets. The results show that different versions of our proposed lightweight student architecture perform with similar accuracy levels compared with the teacher network, while using significantly fewer parameters.